1# RUN: SUPPORT_LIB=%mlir_runner_utils_dir/libmlir_c_runner_utils%shlibext \
2# RUN:   %PYTHON %s | FileCheck %s
3
4import ctypes
5import numpy as np
6import os
7import sys
8
9from mlir import ir
10from mlir import runtime as rt
11
12from mlir.dialects import sparse_tensor as st
13from mlir.dialects import builtin
14from mlir.dialects import func
15from mlir.dialects.linalg.opdsl import lang as dsl
16
17_SCRIPT_PATH = os.path.dirname(os.path.abspath(__file__))
18sys.path.append(_SCRIPT_PATH)
19from tools import sparse_compiler
20
21@dsl.linalg_structured_op
22def sddmm_dsl(
23    A=dsl.TensorDef(dsl.T, dsl.S.M, dsl.S.K),
24    B=dsl.TensorDef(dsl.T, dsl.S.K, dsl.S.N),
25    S=dsl.TensorDef(dsl.T, dsl.S.M, dsl.S.N),
26    C=dsl.TensorDef(dsl.T, dsl.S.M, dsl.S.N, output=True)):
27    C[dsl.D.m,
28      dsl.D.n] += S[dsl.D.m, dsl.D.n] * A[dsl.D.m, dsl.D.k] * B[dsl.D.k, dsl.D.n]
29
30
31def build_SDDMM(attr: st.EncodingAttr):
32    """Build SDDMM kernel.
33
34  This method generates a linalg op with for matrix multiplication using
35  just the Python API. Effectively, a generic linalg op is constructed
36  that computes C(i,j) += S(i,j) SUM_k A(i,k) B(k,j) for sparse S.
37  """
38    module = ir.Module.create()
39    f64 = ir.F64Type.get()
40    a = ir.RankedTensorType.get([8, 8], f64)
41    b = ir.RankedTensorType.get([8, 8], f64)
42    c = ir.RankedTensorType.get([8, 8], f64)
43    s = ir.RankedTensorType.get([8, 8], f64, attr)
44    arguments = [a, b, s, c]
45    with ir.InsertionPoint(module.body):
46
47        @func.FuncOp.from_py_func(*arguments)
48        def sddmm(*args):
49            return sddmm_dsl(args[0], args[1], args[2], outs=[args[3]])
50
51    return module
52
53
54def boilerplate(attr: st.EncodingAttr):
55    """Returns boilerplate code for main driver."""
56    return f"""
57func.func @main(%a: tensor<8x8xf64>,
58           %b: tensor<8x8xf64>,
59           %c: tensor<8x8xf64>) -> tensor<8x8xf64> attributes {{ llvm.emit_c_interface }} {{
60  %t = arith.constant sparse<[[0,0], [0,2], [4,1]], [1.0, 2.0, 3.0]> : tensor<8x8xf64>
61  %s = sparse_tensor.convert %t : tensor<8x8xf64> to tensor<8x8xf64, {attr}>
62  %0 = call @sddmm(%a, %b, %s, %c) : (tensor<8x8xf64>,
63                                      tensor<8x8xf64>,
64                                      tensor<8x8xf64, {attr}>,
65                                      tensor<8x8xf64>) -> tensor<8x8xf64>
66  return %0 : tensor<8x8xf64>
67}}
68"""
69
70
71def build_compile_and_run_SDDMMM(attr: st.EncodingAttr, compiler):
72    # Build.
73    module = build_SDDMM(attr)
74    func = str(module.operation.regions[0].blocks[0].operations[0].operation)
75    module = ir.Module.parse(func + boilerplate(attr))
76
77    # Compile.
78    engine = compiler.compile_and_jit(module)
79
80    # Set up numpy input and buffer for output.
81    a = np.array([[1.1, 2.1, 3.1, 4.1, 5.1, 6.1, 7.1, 8.1],
82                  [1.2, 2.2, 3.2, 4.2, 5.2, 6.2, 7.2, 8.2],
83                  [1.3, 2.3, 3.3, 4.3, 5.3, 6.3, 7.3, 8.3],
84                  [1.4, 2.4, 3.4, 4.4, 5.4, 6.4, 7.4, 8.4],
85                  [1.5, 2.5, 3.5, 4.5, 5.5, 6.5, 7.5, 8.5],
86                  [1.6, 2.6, 3.6, 4.6, 5.6, 6.6, 7.6, 8.6],
87                  [1.7, 2.7, 3.7, 4.7, 5.7, 6.7, 7.7, 8.7],
88                  [1.8, 2.8, 3.8, 4.8, 5.8, 6.8, 7.8, 8.8]], np.float64)
89    b = np.ones((8, 8), np.float64)
90    c = np.zeros((8, 8), np.float64)
91
92    mem_a = ctypes.pointer(ctypes.pointer(rt.get_ranked_memref_descriptor(a)))
93    mem_b = ctypes.pointer(ctypes.pointer(rt.get_ranked_memref_descriptor(b)))
94    mem_c = ctypes.pointer(ctypes.pointer(rt.get_ranked_memref_descriptor(c)))
95
96    # Allocate a MemRefDescriptor to receive the output tensor.
97    # The buffer itself is allocated inside the MLIR code generation.
98    ref_out = rt.make_nd_memref_descriptor(2, ctypes.c_double)()
99    mem_out = ctypes.pointer(ctypes.pointer(ref_out))
100
101    # Invoke the kernel and get numpy output.
102    # Built-in bufferization uses in-out buffers.
103    # TODO: replace with inplace comprehensive bufferization.
104    engine.invoke('main', mem_out, mem_a, mem_b, mem_c)
105
106    # Sanity check on computed result. Only a few elements
107    # are sampled from the full dense matrix multiplication.
108    full_matmul = np.matmul(a, b)
109    expected = np.zeros((8, 8), np.float64)
110    expected[0, 0] = 1.0 * full_matmul[0, 0]
111    expected[0, 2] = 2.0 * full_matmul[0, 2]
112    expected[4, 1] = 3.0 * full_matmul[4, 1]
113    c = rt.ranked_memref_to_numpy(mem_out[0])
114    if np.allclose(c, expected):
115        pass
116    else:
117        quit(f'FAILURE')
118
119
120def main():
121    support_lib = os.getenv('SUPPORT_LIB')
122    assert support_lib is not None, 'SUPPORT_LIB is undefined'
123    if not os.path.exists(support_lib):
124        raise FileNotFoundError(errno.ENOENT, os.strerror(errno.ENOENT),
125                                support_lib)
126
127    # CHECK-LABEL: TEST: testSDDMMM
128    print('\nTEST: testSDDMMM')
129    with ir.Context() as ctx, ir.Location.unknown():
130        count = 0
131        # Loop over various ways to compile and annotate the SDDMM kernel with
132        # a *single* sparse tensor. Note that we deliberate do not exhaustively
133        # search the full state space to reduce runtime of the test. It is
134        # straightforward to adapt the code below to explore more combinations.
135        levels = [[st.DimLevelType.dense, st.DimLevelType.dense],
136                  [st.DimLevelType.dense, st.DimLevelType.compressed],
137                  [st.DimLevelType.compressed, st.DimLevelType.dense],
138                  [st.DimLevelType.compressed, st.DimLevelType.compressed]]
139        orderings = [
140            ir.AffineMap.get_permutation([0, 1]),
141            ir.AffineMap.get_permutation([1, 0])
142        ]
143        for level in levels:
144            for ordering in orderings:
145                for pwidth in [32]:
146                    for iwidth in [32]:
147                        for par in [0]:
148                            for vec in [0, 1]:
149                                for e in [True]:
150                                    vl = 1 if vec == 0 else 16
151                                    attr = st.EncodingAttr.get(level, ordering, pwidth, iwidth)
152                                    opt = (f'parallelization-strategy={par} '
153                                           f'vectorization-strategy={vec} '
154                                           f'vl={vl} enable-simd-index32={e}')
155                                    compiler = sparse_compiler.SparseCompiler(
156                                        options=opt, opt_level=0, shared_libs=[support_lib])
157                                    build_compile_and_run_SDDMMM(attr, compiler)
158                                    count = count + 1
159    # CHECK: Passed 16 tests
160    print('Passed ', count, 'tests')
161
162
163if __name__ == '__main__':
164    main()
165